NIPS 2015

Selected Recap

 

 

 

Roy Hyunjin Han

@crosscompute

Neural networks learn black box models for engineering and industry

  • Training neural networks is an art and a science.
  • Recurrent neural networks learn temporal patterns
  • Reinforcement learning learns control through interactions with the environment.
  • Companies have open sourced software and hardware to train more deep learning experts.

 

(Press DOWN to explore or RIGHT to skip)

Training neural networks

Regularization techniques prevent overfitting (excelling at test examples but failing on new examples).

 

  • Adversarial (fool the model that is being trained)
  • Noise-based (dropout, gradient noise)
  • Miscellaneous (batch normalization, gradient clipping)

 

Leon Gatys et al. Neural Algorithm of Artistic Style. 2015

Recurrent neural networks learn temporal patterns

  • RNNs generalize and surpass HMMs at learning sequences (e.g. speech recognition)
  • LSTM and GRU architectures incorporate a concept of memory (deciding what to remember and forget)
  • Might be useful for early warning systems

Deep reinforcement learning is interactive

Control is learning how to interact with the environment to reach goals (e.g. robotics).

 

Volodymyr Mnih et al. Human-level Control through Deep Reinforcement LearningNature 518, 2015.

Volodymyr Mnih et al. Playing Atari with Deep Reinforcement LearningNIPS Workshop, 2013.

Open source deep learning

 

Lukasz Kaiser, Ilya Sutskever. Neural GPUs learn algorithms2015.

Bayesian inference learns interpretable models for health, science, policy

  • Sometimes it is important to be able to interpret and understand the rules learned by our model.
  • Probabilistic programming is an executable graphical model that performs inference through simulations.
  • Bayesian nonparametrics are an expressive alternative to deep learning.

 

(Press DOWN to explore or RIGHT to skip)

Interpretability is important for health, science, policy

Rich Caruana. Accuracy on the test set is not enough: The risk of deploying unintelligible models in healthcare. NIPS Workshop, 2015.

  • In healthcare, a model can learn dangerous rules that must be screened by an expert. For these cases, purely predictive black box approaches (e.g. deep learning) are unsuitable because it is important to be able to understand what the model learned.

Robert Tibshirani. Some Recent Advances in Post-selection inference. Breiman Invited lecture, NIPS 2015.

Probabilistic programming performs inference through simulations

Bayesian nonparametrics

Non-parametric Bayesian methods are an alternative to deep learning that generates models that are amenable to interpretation, which could prove useful in science.

 

MCMC estimates integrals over distributions. Metropolis-Hastings is popular for sampling from high-dimensional distributions.

 

Zoubin Ghahramani. "Probabilistic machine learning and artificial intelligence." Nature 521, 2015.